Dear everyone,
The approach suggested by Cyril is definitely interesting, however it might turn out to be problematic if voxels are misclassified consistently as GM. In the worst case you lose the "true" thalamus voxels when going with a more conservative threshold, which might be associated with lower GM values compared to cortical GM due to different tissue properties, and end up with just the "false-postive" ones in case they have large(r) GM values.
It might be more promising to start with CSF images in native space, apply some smoothing, threshold the image (thus mimicking a "growth" of the original tissue classified as CSF) and mask the raw GM files with their thresholded, individual "CSF masks". Then apply normalisation and modulation, plus smoothing if necessary. This way you make sure that you get rid of these boundary voxels (be it GM or not). In contrast, if you mask the data after smoothing/during model estimation then the remaining voxels might be "contaminated" by the boundary voxels to some extent nonetheless.
However, this is certainly not perfect either, as you might lose more "true" thalamus voxels in subjects in which there's a larger surface between thalamus and CSF. To make absolutely sure about the results it would probably be necessary to e.g. generate a 3D model of the thalamus.
Best
Helmut
|